The LLM Wiki (Kehle) — assessed against the RDCO vault

Why in vault: founder shared 2026-07-21 with "anything worth stealing / already doing / avoid?" — audit delivered same night. His system (454 pages, 445k words, 7,388 wikilinks, 3 months) is structurally our vault at ~1/7th scale (ours: 3,423 files, 8.7M words, 14,830 wikilinks, git-tracked) with better plumbing in exactly one layer.

His system, compressed

Two-layer ownership (raw/ = human sources, wiki/ = agent-owned compiled understanding; one-way flow) · organize-by-use folders · hierarchical indexes (root ≤25 lines, ~3 reads to anything) · one page template, 14 allowed types, linter-enforced · rules live in files (AGENTS.md constitution, CLAUDE.md = one-line shim) · closed set of 6 named operations (ingest/update/query/lint/enrich/reorganize) each with scar-derived guardrails · deterministic gate: 1,671-line stdlib-Python linter, 20 checks, ~2s, pre-commit hook, warnings blocking · division of labor: script = mechanical, model = semantic, human = taste · light daily (5 min) + heavy weekly (dated report w/ deltas + regression flags) · learnings: healthy KBs sometimes SHRINK; duplication fixed by ownership-not-merging; every rule traces to a failure; move deterministic out of the model; scorecard (15 cold questions, single-clear vs scattered) beats graph views. Still broken for him: retrieval-quality verification, tag sprawl (662 tags — "tags are the wrong tool once wikilinks+indexes exist"), staleness cost forever, writing outran reading (0.68 sources/page).

STEALING (queued as board task 2026-07-21, "Vault gets a deterministic linter")

1. Deterministic linter + commit gate — our nightly /vault-health burns an LLM session on work a 2-second script does better and more repeatably. Convergent with our own 7/8 brigade eval finding ("deterministic axes run as code, not critic-agent discretion"). LLM keeps only semantic judgment.
2. The not-found floor — never claim the vault lacks something until full-text search is empty. Scar from the same day: the after-automation qmd miss (first search missed an existing note; deeper pass found it).
3. 15-question cold scorecard — quarterly; score single-clear vs scattered. His finding that taste externalizes last is worth testing against ours.
4. Consolidation-pass mindset — our vault has never shrunk; 256 stale project docs + dormant newsletter/investing folders (founder archive ruling still pending from the 7/21 health cron).

ALREADY DOING

Agents-as-maintainers (Ray + 5 nightly/weekly crons) · typed frontmatter + confidence tags + provenance/no-invention · untrusted-source quarantine at ingestion (ahead of him — his "reading list is an attack surface" = our standing discipline + injection-park rules) · git · scar-derived rulebook philosophy (CLAUDE.md history = his "changelog of things going wrong") · weekly self-review · one-owner instincts (state files vs vault layering).

SKIP / AVOID

• His hard raw-vs-wiki split: right for personal notes; churn for us (reference notes carry source+assessment together; our real split is vault-vs-state-files).
• Tag skepticism: moot — we never leaned on tags (qmd carries retrieval).
• His no-embeddings-under-10k bet: wrong at our scale; qmd (BM25+vector+hyde) already earns its keep at 3,400+ docs.

Cross-links

[[2026-07-21-huber-12-factor-companies]] (his #11 store-and-learn-from-traces = Kehle's whole system) · feedback_workflow_agent_output_integrity (our scar-ledger equivalent) · the vault-linter board task · /vault-health skill (slimming target).

Why this is in the vault

• Closest public analogue to the RDCO vault's own architecture — 3 months of agent-maintained operation at 1/7th scale, with independently converged design decisions (typed frontmatter, scar-derived rules, one-owner instincts, git-tracked)
• The deterministic linter steal is concrete and actionable: replaces an LLM session with a 2-second Python script for mechanical checks — queued to the board same night
• "Every rule traces to a failure" and "the changelog IS the scar ledger" are independent validation of RDCO's CLAUDE.md-as-changelog-of-things-going-wrong philosophy
• The 15-question cold scorecard is a quarterly vault health metric RDCO currently lacks; the finding that "taste externalizes last" is worth testing against RDCO's own vault

Mapping against Ray Data Co

• Vault operations: deterministic linter + commit gate steal replaces /vault-health LLM session cost with a 2-second script; closed set of 6 named operations (ingest/update/query/lint/enrich/reorganize) is a template for formalizing RDCO vault-agent vocabulary
• Harness-engineering: "untrusted-source quarantine at ingestion" = RDCO's standing injection-park rules (ahead of him); RDCO is structurally ahead on scale (3,400+ docs), retrieval (qmd BM25+vector+hyde), and agent diversity
• Agent L4→L5: his production-traces gap maps exactly to the #11 gap in the 12-factor companies note; three independent builders (Kehle, enginoid, Huber) naming the same gap confirms it as the real L5 frontier
• Vault health: "healthy KBs sometimes SHRINK" and the consolidation-pass mindset are actionable for the 256 stale project docs + dormant folder review the founder has pending